Wubble World

Transcription

Wubble World
Wubble World
Daniel Hewlett, Shane Hoversten, Wesley Kerr,
Paul Cohen, Yu-Han Chang
University of Southern California
Information Sciences Institute
Introducing Wubble World
Overview
• Goals
• Background
– Research
– Gaming
• Games
• Learning Methods
• Current Status and Open Questions
Research Goals
• Wubble World is a virtual environment for
learning the semantics of natural language
from players
• The meanings of words situated in the virtual
environment can be grounded in the objects
and actions occurring in it, enabling richer
semantics
• The relationship between a player and his/her
Wubble is modelled after that of a parent and
a language-learning child
Child Language Learning
• Structured language-centric interaction
– Child: “What's that?”
– Parent: “That's a dog!”
– Shared attention, deictic pointing, other factors
contribute to the child quickly learning what 'dog' means
• Passive observation of parents and peers
– Parent 1 holds up a pan, Parent 1 and Parent 2 speak
some words to each other, including several instances
of 'pan'
– Over time, child learns what 'pan' means
– Better than just simple association, because children
understand intentions of others (theory of mind)
• Everything in between
Gaming Background
• Children spend hours playing games like
Club Penguin, Marian’s World, Neopets
• Much of this time is spent generating
language
• Some language is about what's going on
in the virtual world
– Could an agent in the same virtual world
learn from this language?
Basic Game Model
• Internet games with low system requirements,
playable on-demand and for free
• Kid-friendly content
• Each world includes a main “lobby” area where
players chat with one another
• Players earn points through playing mini-games
• Points are used to buy customizations
• That's it!
Problems for Language
Learning
• Much of the language in such games does not
make direct reference to the goings-on in the
virtual world
– Lots of social chatting
• “I the boss”
• “No, I the boss!”
• “You can be co-boss…”
• Mini-games don't offer much help, as they
rarely require language or involve other
players
Generating Situated Language
• Two Methods
1) Create mini-games explicitly involving language
• Language is player-computer
• ’Blocks Room'
2) Create multi-player games that require
communication about the environment
• Language is player-player
• 'Sheep' Game (upcoming)
●
'Wictionary' game is an attempt to be both
Blocks Room
• 'Blocks World' in a cozy cottage setting
• Only the Wubble can interact with the room
– Walk, jump, look around
– Pick up and put down objects
• The child directs the Wubble using English
NLU Pipeline I
Utterance
Parse Tree
“now go in front of the big green cylinder, please”
Stanford Parser
(ROOT (S
(ADVP (RB now))
(VP (VB go)
(PP (IN in)
(NP
(NP (NN front))
(PP (IN of)
(NP (DT the) (JJ big) (JJ green) (NN cylinder)))))
(, ,) (VB please))))
Tree Transformation
Logical Form
(imperative (subj self)
(verb go)
(prep-phrase (prep in-front-of)
(obj (adj big green)
(noun cylinder))))
NLU Pipeline II
• The referent of each noun phrase is selected
from among the objects in the room
– With the player's help, if necessary
• Prepositions are resolved to a concrete spatial
relation
– Again, possibly with the player's help
• With the meaning of these words determined,
the appropriate action can be executed
Learning in Wubble Room
• Learning is online, human-in-the-loop
• Goal is to learn the meaning of words the child
uses
– Currently: Basic nouns, adjectives, prepositions
– Future: Verbs, composite objects, and more
• Objects are represented as vectors of discrete
features
– Continuous features such as color or size are ‘binned’
• Learning a word means learning the concept that
word denotes
– probability distributions over feature vectors
The Meanings of Words
• The concept associated with the noun
'cube' would be represented as follows:
Weights
COLOR:
Feature Values
Features
SHAPE:
SIZE:
Low-Entropy Weight
Distribution
High-Entropy
Weight
Distribution
Imagination
• Each feature vector has a corresponding
probability distribution (derived from the
weights on feature values)
• A prototypical object is constructed from
sampling these distributions
• All the objects in the room are compared
against this prototype, and the most similar
is the referent
Feedback
• When uncertain, the Wubble will ask the child to
point out the object (and/or location)
• Negative Feedback
– The Wubble responds by selecting a different object
from the room, or querying the player if credit
assignment is ambiguous
• Positive Feedback
– The weights associated with all of the features of the
correct referent are strengthened, for each noun and
adjective of the corresponding noun phrase
– Note: No feedback is implicit positive feedback
Avoiding Credit Assignment
• Dialogue:
–
–
–
–
–
–
–
–
–
–
Player: “go to the right of the cube”
(Wubble walks to the right of a cylinder)
Player: “no”
Wubble: “Is this the cube?” (Points to the cylinder)
Player: “no”
Wubble: “Am I to the right of this?” (Points to the cylinder)
Player: “yes”
Wubble: “Show me the cube.”
(Player points to the cube)
(Wubble walks to the right of the cube)
Updating Weights
• Weights are updated so as to minimize regret
– Method developed by P. Auer, et al.
– Consider the feature values of the correct object
, where
is a value of feature
– The weights corresponding to each value in each
feature vector of the concept is modified
according to the following formula (gamma is the
reward value):
Qualitative Results
• For most nouns, adjectives, and simple
prepositions, a few exposures is
enough to produce correct behavior
– Because player helps the Wubble when it
is uncertain, random exploration is
minimized
• The assumption of independence
between features leads to an inability to
precisely learn certain concepts
Quantitative Results
For evaluation of learning performance, see:
Learning in Wubble World
Wesley Kerr, Shane Hoversten, Daniel Hewlett, Paul
Cohen, Yu-Han Chang
to appear this July in
Proceedings of the International Conference on
Development and Learning (ICDL) 2007
Limitations
• In order for the game to be engaging, the
Wubble has to be able to understand (with some
help) a good portion of the language the child
uses
• This means that the language is constrained to
certain types of relatively simple sentences
• Wubble’s knowledge limits complexity of actions
• What if we wanted to learn from open-ended
language and activity?
Sentence-Scene Corpus
• By observing players in the course of playing a
game and talking about it, a parallel corpus of
sentences and scenes can be constructed
• Ideally
– 'Sentences' are actually parse trees or logical forms
of a wide variety of interesting utterances
– 'Scenes' have structured representations at various
levels of detail (objects, events, etc.) of a dynamic
world
• With such a corpus, it should be possible to
automatically learn associations between the
components of the two structures
Attempt 1: Wictionary
• Players take turns choosing a word or
phrase and drawing it on the screen
• Other players must guess the phrase
• Points are awarded to artist and first
player who guesses right
Pros
• User-generated content and language means
that a relatively simple game can generate a
wide variety of structures and labels for them
– Games like 'Second Life' support content creation
in 3D
• Wrong answers might be interesting as well
– If the right answer was 'duck' and several people
guess 'bird', perhaps a relation exists between
these words
Cons
• Players stick to short phrases, mostly nouns
naming objects
– 'cow', 'person', 'cactus', etc.
• Though the drawing boards supports some
simple dynamics, this feature was rarely
utilized
• The need for other players to guess the
phrase keeps phrases short
• Ultimately, a different direction was needed to
generate more sophisticated situated language
Attempt 2: Sheep
• Team-based multi-player game
Sheep Game
• Team members need to communicate to
coordinate their strategy to herd sheep
toward their side of the field
– Voice communication enables real-time
language during engaging activity
• Offline speech recognition for language processing
– Important game elements require multiple
players working together to utilize effectively
– Player-to-player communication means
language is not constrained
http://www.wubble-world.com